| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501 |
| 4 | """Dots-path speech encoder for inference only (single GPU). |
| 5 | |
| 6 | Ported from cybertron_alm ``dots_audio_encoder/modeling_whisper.py``. |
| 7 | Upstream ``WhisperEncoder`` is exposed as :class:`DotsSpeechEncoder`. |
| 8 | """ |
| 9 | |
| 10 | import math |
| 11 | from typing import Any, ClassVar |
| 12 | |
| 13 | import torch |
| 14 | from torch import nn |
| 15 | from transformers.activations import ACT2FN |
| 16 | from transformers.modeling_outputs import BaseModelOutput |
| 17 | from transformers.modeling_utils import PreTrainedModel |
| 18 | from transformers.models.whisper.configuration_whisper import WhisperConfig |
| 19 | from transformers.utils import logging |
| 20 | |
| 21 | from vllm.vllm_flash_attn import flash_attn_varlen_func, is_fa_version_supported |
| 22 | |
| 23 | _FLASH_ATTN_VERSION = 3 if is_fa_version_supported(3) else 2 |
| 24 | |
| 25 | logger = logging.get_logger(__name__) |
| 26 | |
| 27 | __all__ = ["DotsSpeechEncoder"] |
| 28 | |
| 29 | |
| 30 | class RMSNorm(nn.Module): |
| 31 | def __init__(self, dim: int, eps: float = 1e-6): |
| 32 | super().__init__() |
| 33 | self.weight = nn.Parameter(torch.ones(dim)) |
| 34 | self.eps = eps |
| 35 | |
| 36 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 37 | var = x.pow(2).mean(-1, keepdim=True) |
| 38 | x = x * torch.rsqrt(var + self.eps) |
| 39 | return self.weight * x |
| 40 | |
| 41 | |
| 42 | def swiglu(x: torch.Tensor) -> torch.Tensor: |
| 43 | x1, x2 = x.chunk(2, dim=-1) |
| 44 | return torch.nn.functional.silu(x1) * x2 |
| 45 | |
| 46 | |
| 47 | class RotaryEmbedding(nn.Module): |
| 48 | def __init__( |
| 49 | self, |
| 50 | head_dim: int, |
| 51 | rope_parameters: dict, |
| 52 | base_seq_len: int = 0, |
| 53 | ): |
| 54 | super().__init__() |
| 55 | self.partial_rotary_factor = float( |
| 56 | rope_parameters.get("partial_rotary_factor", 1.0) |
| 57 | ) |
| 58 | self.rope_theta = float(rope_parameters.get("rope_theta", 10000.0)) |
| 59 | self.rope_type = rope_parameters.get("rope_type", "default") |
| 60 | rotary_dim = int(head_dim * self.partial_rotary_factor) |
| 61 | # keep even rotary dimension to align cos/sin pairs |
| 62 | self.rotary_dim = (rotary_dim // 2) * 2 |
| 63 | self.attention_scaling = 1.0 |
| 64 | if self.rope_type != "default": |
| 65 | logger.warning( |
| 66 | "RoPE type %s is not implemented in WhisperLv3; using default", |
| 67 | self.rope_type, |
| 68 | ) |
| 69 | if self.rotary_dim > 0: |
| 70 | inv_freq = 1.0 / ( |
| 71 | self.rope_theta |
| 72 | ** ( |
| 73 | torch.arange(0, self.rotary_dim, 2, dtype=torch.float) |
| 74 | / max(self.rotary_dim, 1) |
| 75 | ) |
| 76 | ) |
| 77 | else: |
| 78 | inv_freq = torch.tensor([]) |
| 79 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 80 | self.base_seq_len = max(0, int(base_seq_len)) |
| 81 | self._cache: ( |
| 82 | tuple[int, torch.dtype, torch.device, torch.Tensor, torch.Tensor] | None |
| 83 | ) = None |
| 84 | |
| 85 | @torch.no_grad() |
| 86 | def get_cos_sin( |
| 87 | self, position_ids: torch.Tensor, dtype: torch.dtype, device: torch.device |
| 88 | ) -> tuple[torch.Tensor | None, torch.Tensor | None]: |
| 89 | if self.rotary_dim == 0: |
| 90 | return None, None |
| 91 | seq_len = position_ids.shape[-1] |
| 92 | if position_ids.shape[0] == 1 and self._cache is not None: |
| 93 | cached_seq_len, cached_dtype, cached_device, cached_cos, cached_sin = ( |
| 94 | self._cache |
| 95 | ) |
| 96 | if ( |
| 97 | cached_seq_len >= seq_len |
| 98 | and cached_dtype == dtype |
| 99 | and cached_device == device |
| 100 | ): |
| 101 | return cached_cos[:, :seq_len, :], cached_sin[:, :seq_len, :] |
| 102 | |
| 103 | if position_ids.shape[0] == 1: |
| 104 | cache_seq_len = max(seq_len, self.base_seq_len) |
| 105 | position_ids = torch.arange(cache_seq_len, device=device)[None, :] |
| 106 | # [1, D/2, 1] |
| 107 | if self.inv_freq.dtype != torch.float32: |
| 108 | inv_freq = 1.0 / ( |
| 109 | self.rope_theta |
| 110 | ** ( |
| 111 | torch.arange( |
| 112 | 0, self.rotary_dim, 2, dtype=torch.float, device=device |
| 113 | ) |
| 114 | / max(self.rotary_dim, 1) |
| 115 | ) |
| 116 | ) |
| 117 | else: |
| 118 | inv_freq = self.inv_freq.to(device=device) |
| 119 | inv_freq_expanded = inv_freq[None, :, None] |
| 120 | # [B, 1, T] |
| 121 | position_ids_expanded = position_ids[:, None, :].float() |
| 122 | # Force float32 since bfloat16 loses precision on long contexts |
| 123 | # See https://github.com/huggingface/transformers/pull/29285 |
| 124 | device_type = device.type if isinstance(device, torch.device) else "cpu" |
| 125 | with torch.autocast(device_type=device_type, enabled=False): |
| 126 | freqs = ( |
| 127 | inv_freq_expanded.float() @ position_ids_expanded.float() |
| 128 | ).transpose(1, 2) |
| 129 | emb = torch.cat((freqs, freqs), dim=-1) |
| 130 | cos = emb.cos() * self.attention_scaling |
| 131 | sin = emb.sin() * self.attention_scaling |
| 132 | cos = cos.to(dtype) |
| 133 | sin = sin.to(dtype) |
| 134 | if position_ids.shape[0] == 1: |
| 135 | self._cache = (position_ids.shape[-1], dtype, device, cos, sin) |
| 136 | return cos[:, :seq_len, :], sin[:, :seq_len, :] |
| 137 | return cos, sin |
| 138 | |
| 139 | |
| 140 | def rotate_half(x: torch.Tensor) -> torch.Tensor: |
| 141 | x1 = x[..., : x.shape[-1] // 2] |
| 142 | x2 = x[..., x.shape[-1] // 2 :] |
| 143 | return torch.cat((-x2, x1), dim=-1) |
| 144 | |
| 145 | |
| 146 | def apply_rotary_pos_emb( |
| 147 | q: torch.Tensor, |
| 148 | k: torch.Tensor, |
| 149 | cos: torch.Tensor | None, |
| 150 | sin: torch.Tensor | None, |
| 151 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 152 | if cos is None or sin is None: |
| 153 | return q, k |
| 154 | |
| 155 | if q.dim() == 4: |
| 156 | # [B, T, H, D] |
| 157 | rotary_dim = cos.shape[-1] |
| 158 | cos = cos.unsqueeze(2) # [B, T, 1, rotary_dim] |
| 159 | sin = sin.unsqueeze(2) |
| 160 | q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] |
| 161 | k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] |
| 162 | q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin) |
| 163 | k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin) |
| 164 | return torch.cat([q_embed, q_pass], dim=-1), torch.cat( |
| 165 | [k_embed, k_pass], dim=-1 |
| 166 | ) |
| 167 | |
| 168 | if q.dim() == 3: |
| 169 | # [tokens, H, D] (varlen flattened) |
| 170 | rotary_dim = cos.shape[-1] |
| 171 | cos = cos.unsqueeze(1) # [tokens, 1, rotary_dim] |
| 172 | sin = sin.unsqueeze(1) |
| 173 | q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] |
| 174 | k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] |
| 175 | q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin) |
| 176 | k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin) |
| 177 | return torch.cat([q_embed, q_pass], dim=-1), torch.cat( |
| 178 | [k_embed, k_pass], dim=-1 |
| 179 | ) |
| 180 | |
| 181 | return q, k |
| 182 | |
| 183 | |
| 184 | class WhisperPositionalEmbedding(nn.Embedding): |
| 185 | def __init__( |
| 186 | self, num_positions: int, embedding_dim: int, padding_idx: int | None = None |
| 187 | ): |
| 188 | super().__init__(num_positions, embedding_dim) |
| 189 | |
| 190 | def forward(self, input_ids, past_key_values_length=0): |
| 191 | return self.weight[ |
| 192 | past_key_values_length : past_key_values_length + input_ids.shape[1] |
| 193 | ] |
| 194 | |
| 195 | |
| 196 | class WhisperAttention(nn.Module): |
| 197 | """Multi-headed attention from 'Attention Is All You Need' paper.""" |
| 198 | |
| 199 | def __init__( |
| 200 | self, |
| 201 | embed_dim: int, |
| 202 | num_heads: int, |
| 203 | dropout: float = 0.0, |
| 204 | is_decoder: bool = False, |
| 205 | bias: bool = True, |
| 206 | ): |
| 207 | super().__init__() |
| 208 | self.embed_dim = embed_dim |
| 209 | self.num_heads = num_heads |
| 210 | self.dropout = dropout |
| 211 | self.head_dim = embed_dim // num_heads |
| 212 | |
| 213 | if (self.head_dim * num_heads) != self.embed_dim: |
| 214 | raise ValueError( |
| 215 | f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" |
| 216 | f" and `num_heads`: {num_heads})." |
| 217 | ) |
| 218 | self.scaling = self.head_dim**-0.5 |
| 219 | self.is_decoder = is_decoder |
| 220 | |
| 221 | self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False) |
| 222 | self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 223 | self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 224 | self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) |
| 225 | |
| 226 | def forward_flash_attn( |
| 227 | self, |
| 228 | hidden_states, |
| 229 | cu_seqlens_q=None, |
| 230 | cu_seqlens_kv=None, |
| 231 | max_seqlen_q=None, |
| 232 | max_seqlen_kv=None, |
| 233 | output_attentions=False, |
| 234 | rotary_cos=None, |
| 235 | rotary_sin=None, |
| 236 | ): |
| 237 | """Dense eager attention with SGLang FA3 for packed variable-length input.""" |
| 238 | query_states = self.q_proj(hidden_states) |
| 239 | key_states = self.k_proj(hidden_states) |
| 240 | value_states = self.v_proj(hidden_states) |
| 241 | if cu_seqlens_q is None: |
| 242 | bsz, tgt_len, _ = hidden_states.size() |
| 243 | query_states = query_states.view( |
| 244 | bsz, tgt_len, self.num_heads, self.head_dim |
| 245 | ) |
| 246 | key_states = key_states.view(bsz, tgt_len, self.num_heads, self.head_dim) |
| 247 | value_states = value_states.view( |
| 248 | bsz, tgt_len, self.num_heads, self.head_dim |
| 249 | ) |
| 250 | if rotary_cos is not None and rotary_sin is not None: |
| 251 | cos = rotary_cos[:, :tgt_len, :] |
| 252 | sin = rotary_sin[:, :tgt_len, :] |
| 253 | query_states, key_states = apply_rotary_pos_emb( |
| 254 | query_states, key_states, cos, sin |
| 255 | ) |
| 256 | attn_output, attn_probs = self._eager_attention( |
| 257 | query_states, key_states, value_states, output_attentions |
| 258 | ) |
| 259 | attn_output = attn_output.view(bsz, tgt_len, self.embed_dim) |
| 260 | else: |
| 261 | query_states = query_states.view(-1, self.num_heads, self.head_dim) |
| 262 | key_states = key_states.view(-1, self.num_heads, self.head_dim) |
| 263 | value_states = value_states.view(-1, self.num_heads, self.head_dim) |
| 264 | if rotary_cos is not None and rotary_sin is not None: |
| 265 | query_states, key_states = apply_rotary_pos_emb( |
| 266 | query_states, key_states, rotary_cos, rotary_sin |
| 267 | ) |
| 268 | result = flash_attn_varlen_func( |
| 269 | query_states, |
| 270 | key_states, |
| 271 | value_states, |
| 272 | cu_seqlens_q=cu_seqlens_q, |
| 273 | cu_seqlens_k=cu_seqlens_kv, |
| 274 | max_seqlen_q=max_seqlen_q, |
| 275 | max_seqlen_k=max_seqlen_kv, |
| 276 | softmax_scale=self.scaling, |
| 277 | causal=self.is_decoder, |
| 278 | return_softmax_lse=output_attentions, |
| 279 | fa_version=_FLASH_ATTN_VERSION, |
| 280 | ) |
| 281 | if isinstance(result, tuple): |
| 282 | attn_output = result[0] |
| 283 | attn_probs = result[1] if output_attentions else None |
| 284 | else: |
| 285 | attn_output = result |
| 286 | attn_probs = None |
| 287 | attn_output = attn_output.view(-1, self.embed_dim) |
| 288 | |
| 289 | return self.out_proj(attn_output), attn_probs |
| 290 | |
| 291 | def _eager_attention(self, query, key, value, output_attentions): |
| 292 | query = query.transpose(1, 2) |
| 293 | key = key.transpose(1, 2) |
| 294 | value = value.transpose(1, 2) |
| 295 | scores = torch.matmul(query, key.transpose(-2, -1)) * self.scaling |
| 296 | if self.is_decoder: |
| 297 | seq_len = scores.shape[-1] |
| 298 | causal_mask = torch.ones( |
| 299 | seq_len, seq_len, dtype=torch.bool, device=scores.device |
| 300 | ).triu(1) |
| 301 | scores.masked_fill_(causal_mask, torch.finfo(scores.dtype).min) |
| 302 | probabilities = torch.softmax(scores, dim=-1, dtype=torch.float32).to( |
| 303 | query.dtype |
| 304 | ) |
| 305 | output = torch.matmul(probabilities, value).transpose(1, 2) |
| 306 | return output, probabilities if output_attentions else None |
| 307 | |
| 308 | |
| 309 | # Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Whisper |
| 310 | class WhisperEncoderLayer(nn.Module): |
| 311 | def __init__(self, config: WhisperConfig): |
| 312 | super().__init__() |
| 313 | self.embed_dim = config.d_model |
| 314 | use_causal = getattr(config, "use_causal", False) |
| 315 | self.self_attn = WhisperAttention( |
| 316 | embed_dim=self.embed_dim, |
| 317 | num_heads=config.encoder_attention_heads, |
| 318 | dropout=config.attention_dropout, |
| 319 | is_decoder=use_causal, # enable causal attention when use_causal=True |
| 320 | ) |
| 321 | self.use_causal = use_causal |
| 322 | norm_cls = RMSNorm if getattr(config, "use_rms_norm", False) else nn.LayerNorm |
| 323 | self.self_attn_layer_norm = norm_cls(self.embed_dim) |
| 324 | self.dropout = config.dropout |
| 325 | self.activation_fn = ( |
| 326 | ACT2FN[config.activation_function] |
| 327 | if config.activation_function != "swiglu" |
| 328 | else swiglu |
| 329 | ) |
| 330 | self.use_swiglu = config.activation_function == "swiglu" |
| 331 | self.activation_dropout = config.activation_dropout |
| 332 | ffn_dim = config.encoder_ffn_dim |
| 333 | fc1_out = ffn_dim * 2 if self.use_swiglu else ffn_dim |
| 334 | self.fc1 = nn.Linear(self.embed_dim, fc1_out) |
| 335 | self.fc2 = nn.Linear(ffn_dim, self.embed_dim) |
| 336 | self.final_layer_norm = norm_cls(self.embed_dim) |
| 337 | |
| 338 | def forward( |
| 339 | self, |
| 340 | hidden_states: torch.Tensor, |
| 341 | cu_seqlens_q: torch.Tensor = None, |
| 342 | cu_seqlens_kv: torch.Tensor = None, |
| 343 | max_seqlen_q: int | None = None, |
| 344 | max_seqlen_kv: int | None = None, |
| 345 | output_attentions: bool = False, |
| 346 | rotary_cos: torch.Tensor | None = None, |
| 347 | rotary_sin: torch.Tensor | None = None, |
| 348 | ) -> tuple[Any, ...]: |
| 349 | """Args: |
| 350 | hidden_states: Input to the layer of shape `(batch, seq_len, |
| 351 | embed_dim)`, or `(total_tokens, embed_dim)` when `cu_seqlens_q` |
| 352 | is given. |
| 353 | cu_seqlens_q: Cumulative query sequence lengths for packed |
| 354 | variable-length input. If `None`, the input is treated as a |
| 355 | padded batch. |
| 356 | cu_seqlens_kv: Cumulative key/value sequence lengths for packed |
| 357 | variable-length input. |
| 358 | max_seqlen_q: Longest query sequence in the packed batch. |
| 359 | max_seqlen_kv: Longest key/value sequence in the packed batch. |
| 360 | output_attentions: Whether to also return the attention weights. |
| 361 | rotary_cos: Cosine component of the rotary position embedding. |
| 362 | rotary_sin: Sine component of the rotary position embedding. |
| 363 | |
| 364 | Returns: |
| 365 | A tuple of the output hidden states, followed by the attention |
| 366 | weights if `output_attentions` is `True`. |
| 367 | |
| 368 | """ |
| 369 | residual = hidden_states |
| 370 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 371 | |
| 372 | hidden_states, attn_weights = self.self_attn.forward_flash_attn( |
| 373 | hidden_states=hidden_states, |
| 374 | cu_seqlens_q=cu_seqlens_q, |
| 375 | cu_seqlens_kv=cu_seqlens_kv, |
| 376 | max_seqlen_q=max_seqlen_q, |
| 377 | max_seqlen_kv=max_seqlen_kv, |
| 378 | output_attentions=output_attentions, |
| 379 | rotary_cos=rotary_cos, |
| 380 | rotary_sin=rotary_sin, |
| 381 | ) |
| 382 | hidden_states = nn.functional.dropout( |
| 383 | hidden_states, p=self.dropout, training=self.training |
| 384 | ) |
| 385 | hidden_states = residual + hidden_states |
| 386 | |
| 387 | residual = hidden_states |
| 388 | hidden_states = self.final_layer_norm(hidden_states) |
| 389 | if self.use_swiglu: |
| 390 | hidden_states = swiglu(self.fc1(hidden_states)) |
| 391 | else: |
| 392 | hidden_states = self.activation_fn(self.fc1(hidden_states)) |
| 393 | hidden_states = nn.functional.dropout( |
| 394 | hidden_states, p=self.activation_dropout, training=self.training |
| 395 | ) |
| 396 | hidden_states = self.fc2(hidden_states) |
| 397 | hidden_states = nn.functional.dropout( |
| 398 | hidden_states, p=self.dropout, training=self.training |
| 399 | ) |
| 400 | hidden_states = residual + hidden_states |
| 401 | |
| 402 | if hidden_states.dtype == torch.float16 and ( |
| 403 | torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any() |
| 404 | ): |
| 405 | clamp_value = torch.finfo(hidden_states.dtype).max - 1000 |
| 406 | hidden_states = torch.clamp( |
| 407 | hidden_states, min=-clamp_value, max=clamp_value |
| 408 | ) |
| 409 | |
| 410 | outputs: tuple[Any, ...] = (hidden_states,) |
| 411 | |
| 412 | if output_attentions: |
| 413 | outputs += (attn_weights,) |
| 414 | |
| 415 | return outputs |
| 416 | |
| 417 | |
| 418 | class DotsSpeechPreTrainedModel(PreTrainedModel): |
| 419 | config_class = WhisperConfig |
| 420 | base_model_prefix = "model" |
| 421 | main_input_name = "input_features" |
| 422 | supports_gradient_checkpointing = False |
| 423 | _no_split_modules: ClassVar[list[str]] = ["WhisperEncoderLayer"] |
| 424 | |
| 425 | def _init_weights(self, module): |
| 426 | std = self.config.init_std |
| 427 | if isinstance(module, (nn.Linear, nn.Conv1d, nn.Conv2d)): |
| 428 | module.weight.data.normal_(mean=0.0, std=std) |
| 429 | if module.bias is not None: |
| 430 | module.bias.data.zero_() |
| 431 | elif isinstance(module, nn.Embedding): |
| 432 | module.weight.data.normal_(mean=0.0, std=std) |
| 433 | if module.padding_idx is not None: |
| 434 | module.weight.data[module.padding_idx].zero_() |
| 435 | |
| 436 | |
| 437 | class DotsSpeechEncoder(DotsSpeechPreTrainedModel): |
| 438 | """Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a |
| 439 | [`WhisperEncoderLayer`]. |
| 440 | |
| 441 | Args: |
| 442 | config: WhisperConfig |
| 443 | |
| 444 | """ |
| 445 | |
| 446 | def __init__(self, config: WhisperConfig): |
| 447 | super().__init__(config) |
| 448 | self.dropout = config.dropout |
| 449 | |
| 450 | embed_dim = config.d_model |
| 451 | self.use_latent_input = getattr(config, "use_latent_input", False) |
| 452 | self.use_causal = getattr(config, "use_causal", False) |
| 453 | self.use_conv2d_stem = getattr(config, "use_conv2d_stem", False) |
| 454 | latent_dim = getattr(config, "latent_dim", None) |
| 455 | if self.use_latent_input and latent_dim is None: |
| 456 | raise ValueError( |
| 457 | "DotsSpeechEncoder: use_latent_input=True requires config.latent_dim" |
| 458 | ) |
| 459 | if self.use_conv2d_stem and self.use_latent_input: |
| 460 | raise ValueError( |
| 461 | "DotsSpeechEncoder: use_conv2d_stem and use_latent_input are mutually exclusive" |
| 462 | ) |
| 463 | self.num_mel_bins = ( |
| 464 | latent_dim if self.use_latent_input and latent_dim is not None else 128 |
| 465 | ) |
| 466 | self.padding_idx = config.pad_token_id |
| 467 | self.max_source_positions = config.max_source_positions |
| 468 | self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0 |
| 469 | |
| 470 | if self.use_conv2d_stem: |
| 471 | # Conv2D stem: 3 layers of stride-2 for 8x downsampling |
| 472 | dhs = getattr(config, "downsample_hidden_size", 480) |
| 473 | # Causal: keep freq padding, remove time padding (handled by pad(14,0) in forward) |
| 474 | conv_padding = (1, 0) if self.use_causal else 1 |
| 475 | self.conv2d1 = nn.Conv2d( |
| 476 | 1, dhs, kernel_size=3, stride=2, padding=conv_padding |
| 477 | ) |
| 478 | self.conv2d2 = nn.Conv2d( |
| 479 | dhs, dhs, kernel_size=3, stride=2, padding=conv_padding |
| 480 | ) |
| 481 | self.conv2d3 = nn.Conv2d( |
| 482 | dhs, dhs, kernel_size=3, stride=2, padding=conv_padding |
| 483 | ) |
| 484 | # After 3x stride-2 on freq=128: 128→64→32→16; linear projects dhs*16 → embed_dim |
| 485 | freq_after = self.num_mel_bins |
| 486 | for _ in range(3): |
| 487 | freq_after = (freq_after + 1) // 2 |
| 488 | self.conv_out = nn.Linear(dhs * freq_after, embed_dim, bias=False) |
| 489 | self.conv1 = None |
| 490 | self.conv2 = None |
| 491 | elif self.use_latent_input: |
| 492 | # Latent path keeps length (stride=1) and applies GLU after conv1 |
| 493 | self.conv1 = nn.Conv1d( |
| 494 | self.num_mel_bins, embed_dim * 2, kernel_size=3, stride=1, padding=1 |
| 495 | ) |
| 496 | self.conv2 = nn.Conv1d( |
| 497 | embed_dim, embed_dim, kernel_size=3, stride=1, padding=1 |
| 498 | ) |
| 499 | else: |
| 500 | self.conv1 = nn.Conv1d( |
| 501 | self.num_mel_bins, embed_dim, kernel_size=3, padding=1 |
| 502 | ) |
| 503 | self.conv2 = nn.Conv1d( |
| 504 | embed_dim, embed_dim, kernel_size=3, stride=2, padding=1 |
| 505 | ) |
| 506 | |
| 507 | self.use_rope = getattr(config, "use_rope", False) |
| 508 | rope_parameters = getattr(config, "rope_parameters", {}) or {} |
| 509 | self.rope_parameters = rope_parameters |
| 510 | if self.use_rope: |
| 511 | head_dim = embed_dim // config.encoder_attention_heads |
| 512 | self.rotary_embedding = RotaryEmbedding( |
| 513 | head_dim, |
| 514 | rope_parameters, |
| 515 | base_seq_len=self.max_source_positions, |
| 516 | ) |
| 517 | self.embed_positions = None |
| 518 | else: |
| 519 | self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim) |
| 520 | |
| 521 | self.layers = nn.ModuleList( |
| 522 | [WhisperEncoderLayer(config) for _ in range(config.encoder_layers)] |
| 523 | ) |
| 524 | norm_cls = RMSNorm if getattr(config, "use_rms_norm", False) else nn.LayerNorm |
| 525 | self.layer_norm = norm_cls(config.d_model) |
| 526 | |
| 527 | self.post_init() |
| 528 | |
| 529 | def get_input_embeddings(self) -> nn.Module: |
| 530 | if self.use_conv2d_stem: |
| 531 | return self.conv2d1 |
| 532 | return self.conv1 |
| 533 | |
| 534 | def set_input_embeddings(self, value: nn.Module): |
| 535 | self.conv1 = value |
| 536 | |
| 537 | @staticmethod |
| 538 | def _temporal_mask(feat, valid_lens): |
| 539 | """Zero out temporal positions >= valid_lens. feat: [B,C,F,T], valid_lens: [B] tensor.""" |
| 540 | T = feat.shape[-1] |
| 541 | mask = torch.arange(T, device=feat.device)[None, :] < valid_lens[:, None] |
| 542 | return mask[:, None, None, :] |
| 543 | |
| 544 | def _conv2d_stem_one_chunk(self, chunk, chunk_valid_mel_lens=None): |
| 545 | """Run 3x Conv2d(stride=2) + GELU with per-layer masking.""" |
| 546 | # Causal: left-pad time by 14 = 2 + 4 + 8 (receptive field of the 3-layer |
| 547 | # stride-2 stack mapped back to input time). Combined with conv padding=(1,0), |
| 548 | # this makes the whole conv2d stem strictly causal. |
| 549 | if self.use_causal: |
| 550 | chunk = nn.functional.pad(chunk, (14, 0)) |
| 551 | if chunk_valid_mel_lens is not None: |
| 552 | chunk_valid_mel_lens = chunk_valid_mel_lens + 14 |
| 553 | # Step 0: mask mel — silence_mel(-1.5) → 0 |
| 554 | if chunk_valid_mel_lens is not None: |
| 555 | chunk = chunk * self._temporal_mask(chunk, chunk_valid_mel_lens) |
| 556 | chunk = nn.functional.gelu(self.conv2d1(chunk)) |
| 557 | # Step 1: mask conv1 output — gelu(conv(0)+bias) → 0 |
| 558 | if chunk_valid_mel_lens is not None: |
| 559 | chunk_valid_mel_lens = (chunk_valid_mel_lens + 1) // 2 |
| 560 | chunk = chunk * self._temporal_mask(chunk, chunk_valid_mel_lens) |
| 561 | chunk = nn.functional.gelu(self.conv2d2(chunk)) |
| 562 | # Step 2: mask conv2 output — gelu(bias) → 0 |
| 563 | if chunk_valid_mel_lens is not None: |
| 564 | chunk_valid_mel_lens = (chunk_valid_mel_lens + 1) // 2 |
| 565 | chunk = chunk * self._temporal_mask(chunk, chunk_valid_mel_lens) |
| 566 | chunk = nn.functional.gelu(self.conv2d3(chunk)) |
| 567 | # Step 3: mask conv3 output — gelu(bias) → 0 |
| 568 | if chunk_valid_mel_lens is not None: |
| 569 | chunk_valid_mel_lens = (chunk_valid_mel_lens + 1) // 2 |
| 570 | chunk = chunk * self._temporal_mask(chunk, chunk_valid_mel_lens) |
| 571 | return chunk |
| 572 | |
| 573 | def _forward_conv2d_stem( |
| 574 | self, input_features, input_seq_lens=None, audio_sample_lens=None |
| 575 | ): |
| 576 | """Conv2D stem: 3x Conv2d(stride=2) → 8x downsample. Returns [B, T/8, embed_dim].""" |
| 577 | x = input_features.unsqueeze(1) # [B, 1, 128, T] |
| 578 | |
| 579 | if audio_sample_lens is not None: |
| 580 | hop_length = 160 |
| 581 | if not isinstance(audio_sample_lens, torch.Tensor): |
| 582 | audio_sample_lens = torch.tensor(audio_sample_lens, device=x.device) |
| 583 | valid_mel_lens = audio_sample_lens.to(x.device) // hop_length |
| 584 | else: |
| 585 | valid_mel_lens = None |
| 586 | x = self._conv2d_stem_one_chunk(x, valid_mel_lens) |
| 587 | # [B, dhs, F_out, T/8] → [B, T/8, dhs*freq_after] → [B, T/8, embed_dim] |
| 588 | batch, channels, frequency, time = x.shape |
| 589 | x = x.permute(0, 3, 1, 2).reshape(batch, time, channels * frequency) |
| 590 | inputs_embeds = self.conv_out(x) |
| 591 | |
| 592 | return inputs_embeds |
| 593 | |
| 594 | def _forward_conv1d_stem(self, input_features): |
| 595 | """Standard Conv1d stem: conv1 + conv2(stride=2), 2x downsample. Returns [B, T/2, embed_dim].""" |
| 596 | assert self.conv1 is not None and self.conv2 is not None |
| 597 | if self.use_causal: |
| 598 | x = nn.functional.pad(input_features, (2, 0)) |
| 599 | x = nn.functional.gelu(self.conv1(x)) |
| 600 | x = nn.functional.pad(x, (2, 0)) |
| 601 | x = nn.functional.gelu(self.conv2(x)) |
| 602 | else: |
| 603 | x = nn.functional.gelu(self.conv1(input_features)) |
| 604 | x = nn.functional.gelu(self.conv2(x)) |
| 605 | return x.permute(0, 2, 1) |
| 606 | |
| 607 | def _forward_latent_stem(self, input_features): |
| 608 | """Latent input stem: conv1+GLU + conv2, stride=1, no downsample. Returns [B, T, embed_dim].""" |
| 609 | assert self.conv1 is not None and self.conv2 is not None |
| 610 | if self.use_causal: |
| 611 | x = nn.functional.pad(input_features, (2, 0)) |
| 612 | x = nn.functional.glu(self.conv1(x), dim=1) |
| 613 | x = nn.functional.pad(x, (2, 0)) |
| 614 | x = nn.functional.gelu(self.conv2(x)) |
| 615 | else: |
| 616 | x = nn.functional.glu(self.conv1(input_features), dim=1) |
| 617 | x = nn.functional.gelu(self.conv2(x)) |
| 618 | return x.permute(0, 2, 1) |
| 619 | |
| 620 | def forward( |
| 621 | self, |
| 622 | input_features, |
| 623 | input_seq_lens=None, |
| 624 | audio_sample_lens=None, |
| 625 | output_attentions=None, |
| 626 | output_hidden_states=None, |
| 627 | return_dict=None, |
| 628 | ): |
| 629 | """Mel `[B, n_mels, T]` → encoder hidden states. Inference-only.""" |
| 630 | output_attentions = ( |
| 631 | output_attentions |
| 632 | if output_attentions is not None |
| 633 | else self.config.output_attentions |
| 634 | ) |
| 635 | output_hidden_states = ( |
| 636 | output_hidden_states |
| 637 | if output_hidden_states is not None |
| 638 | else self.config.output_hidden_states |
| 639 | ) |
| 640 | return_dict = ( |
| 641 | return_dict if return_dict is not None else self.config.use_return_dict |
| 642 | ) |
| 643 | if self.use_conv2d_stem: |
| 644 | inputs_embeds = self._forward_conv2d_stem( |
| 645 | input_features, input_seq_lens, audio_sample_lens |
| 646 | ) |
| 647 | elif self.use_latent_input: |
| 648 | inputs_embeds = self._forward_latent_stem(input_features) |
| 649 | else: |
| 650 | inputs_embeds = self._forward_conv1d_stem(input_features) |
| 651 | rotary_cos = None |
| 652 | rotary_sin = None |
| 653 | position_ids = torch.arange( |
| 654 | inputs_embeds.shape[1], device=inputs_embeds.device |
| 655 | )[None, :] |
| 656 | if self.use_rope: |
| 657 | rotary_cos, rotary_sin = self.rotary_embedding.get_cos_sin( |
| 658 | position_ids, inputs_embeds.dtype, inputs_embeds.device |
| 659 | ) |
| 660 | hidden_states = inputs_embeds |
| 661 | else: |
| 662 | assert self.embed_positions is not None |
| 663 | embed_pos = self.embed_positions.weight[: inputs_embeds.shape[1]] |
| 664 | hidden_states = inputs_embeds + embed_pos |
| 665 | |
| 666 | encoder_states: tuple[torch.Tensor, ...] | None = ( |
| 667 | () if output_hidden_states else None |
| 668 | ) |
| 669 | all_attentions: tuple[torch.Tensor, ...] | None = ( |
| 670 | () if output_attentions else None |
| 671 | ) |
| 672 | |
| 673 | if input_seq_lens is not None: |
| 674 | # Build varlen metadata and pack valid tokens without per-sample cat loops. |
| 675 | # Read max on whatever device the caller provided: when it is a CPU |
| 676 | # tensor this avoids a device->host sync (one per forward). |
| 677 | max_seqlen_q = int(input_seq_lens.max().item()) |
| 678 | input_seq_lens = input_seq_lens.to( |
| 679 | device=hidden_states.device, dtype=torch.long |
| 680 | ) |
| 681 | B, S, D = hidden_states.shape |
| 682 | max_seqlen_kv = max_seqlen_q |
| 683 | cu_seqlens_q = torch.nn.functional.pad( |
| 684 | input_seq_lens.cumsum(0, dtype=torch.int32), (1, 0) |
| 685 | ) |
| 686 | cu_seqlens_kv = cu_seqlens_q |
| 687 | |
| 688 | token_positions = torch.arange(S, device=hidden_states.device)[None, :] |
| 689 | valid_token_mask = token_positions < input_seq_lens[:, None] |
| 690 | hidden_states = hidden_states[valid_token_mask] |
| 691 | if rotary_cos is not None and rotary_sin is not None: |
| 692 | packed_positions = token_positions.expand(B, S)[valid_token_mask] |
| 693 | rotary_cos = rotary_cos.squeeze(0).index_select(0, packed_positions) |
| 694 | rotary_sin = rotary_sin.squeeze(0).index_select(0, packed_positions) |
| 695 | else: |
| 696 | cu_seqlens_q = None |
| 697 | cu_seqlens_kv = None |
| 698 | max_seqlen_q = None |
| 699 | max_seqlen_kv = None |
| 700 | |
| 701 | for encoder_layer in self.layers: |
| 702 | if output_hidden_states: |
| 703 | assert encoder_states is not None |
| 704 | encoder_states = encoder_states + (hidden_states,) |
| 705 | layer_outputs = encoder_layer( |
| 706 | hidden_states, |
| 707 | cu_seqlens_q, |
| 708 | cu_seqlens_kv, |
| 709 | max_seqlen_q, |
| 710 | max_seqlen_kv, |
| 711 | output_attentions=output_attentions, |
| 712 | rotary_cos=rotary_cos, |
| 713 | rotary_sin=rotary_sin, |
| 714 | ) |
| 715 | hidden_states = layer_outputs[0] |
| 716 | if output_attentions: |
| 717 | assert all_attentions is not None |
| 718 | all_attentions = all_attentions + (layer_outputs[1],) |
| 719 | |
| 720 | if input_seq_lens is not None: |
| 721 | # Recover packed varlen output directly on device. |
| 722 | recover_positions = torch.arange(max_seqlen_q, device=hidden_states.device)[ |
| 723 | None, : |
| 724 | ] |
| 725 | recover_mask = recover_positions < input_seq_lens[:, None] |
| 726 | recovered = hidden_states.new_zeros(B, max_seqlen_q, D) |
| 727 | recovered[recover_mask] = hidden_states |
| 728 | hidden_states = recovered |
| 729 | |
| 730 | hidden_states = self.layer_norm(hidden_states) |
| 731 | if output_hidden_states: |
| 732 | assert encoder_states is not None |
| 733 | encoder_states = encoder_states + (hidden_states,) |
| 734 | |
| 735 | if not return_dict: |
| 736 | return tuple( |
| 737 | v |
| 738 | for v in [hidden_states, encoder_states, all_attentions] |
| 739 | if v is not None |
| 740 | ) |
| 741 | return BaseModelOutput( |
| 742 | last_hidden_state=hidden_states, |
| 743 | hidden_states=encoder_states, |
| 744 | attentions=all_attentions, |
| 745 | ) |